The released artifacts consist of FP8 quantized weights and configuration files in Hugging Face Transformers format. Fine-grained FP8 quantization with a block size of 128 has been adopted, and the team states that performance metrics are nearly equivalent to the original model. It is reported to be compatible with inference frameworks such as vLLM, SGLang, and TokenSpeed.
Qwen3.8 is a series that has seen significant performance improvements in coding and long-term agent tasks. The 27B model is designed as a vision-language model capable of understanding images and videos, aiming to complete complex multi-step tasks more reliably.
Hosted versions are planned for the cloud service "Qwen Cloud." It is stated that these will support a 1 million token context by default and that official built-in tools will also be available. The service is expected to launch soon.
Thinking mode is enabled by default. The depth of inference can be adjusted using the reasoning_effort parameter, allowing for cost management. However, it is explained that while lower settings may result in faster responses per turn, they may increase retries overall, leading to higher latency and token consumption.